Guido Borghi, Simone Calderara, Rita Cucchiara, Roberto Vezzani
Università di Modena e Reggio Emilia, Italy
Deep learning has changed the computer vision science in-toto and revolutionized the panorama of techniques and solutions for human detection, people tracking, people pose understanding and human action recognition.
In particular, current research approaches are a mix of data-driven and model driven solutions, by combining the power of large amount of annotated data for supervised learning and top-down solutions based on 2D and 3D model of humans. Depending on the goal of the vision problem, different solutions have been proposed where the model is only an easy starting point, possibly with as low as possible constraints, and data are strongly used for training neural network applications. The definition of human models are well assessed (3D body models, humans body parts and their joints, 3D heads, ..) but the problem of available data is still not solved completely. As well as data available are not enough, data augmentation by synthetic data and data transfer from one modality to other one are the most common research topics.
This tutorial will discuss in detail these points according with specific applications in video-surveillance and driver-vehicle interaction. Several examples on ongoing projects will be presented. More specifically the topics will be covered are:
- Human detection by pose: the state of the art, new solutions in 2D and in 3D coping with body part occlusions;
- Augmenting data for supervised learning by synthetic data: the experience of JGTA Datset at UNIMORE the largest dataset of human joints and pose
- Domain transfer issues from synthetic to real video context for pose understanding and short term tracking
- Pose detection in head and upper body from range and RGB data: experiences and domain transfer. The poseidon architecture and the related Pandora Dataset
Guido Borghi is a post-doc University of Modena and Reggio Emilia working in 3D head pose estimation from depth images. He is held his Phd at AImagelab, UNIMORE, in a project of driver attention anlaysis in vehicle, also spending 6 months in Stanford University. He is working in collaboration with RedVision Lab UNIMORE-FERRARI and is involved in research for anomaly detection with Italian Railway Company (RFI).
Simone Calderara is Associate Professor University of Modena and Reggio Emilia working on machine learning and deep elarning for human and animal behaviour understanding, tracking and video-surveillance. In the past eh worked on crowd analysis,. He is responsible of the UNIMORE Unit of the National project COSMOS on people tracking and co-responsible of the work with Panasonic Beta Lab for tracking humans and motion analysis from vehicle
Roberto Vezzani is Associate Professor University of Modena and Reggio Emilia working on Intelligent IoT, 3D computer vision , responsable of the RedVision Lab UNIMORE-FERRARI for human-vehicle interaction. He worked for a long time in human behaviour understanding for detection on posture, falling situation, tracking using floor sensors and recently working on domain transfer for range and RGB head pose analysis. He is working in H2020 DeepHealth and ColRobot project.
Rita Cucchiara is Full Professor University of Modena and Reggio Emilia working on deep learning and computer vision for video surveillance and automotive applications, and in particular on human pose estimation, action analysis and tracking, as well as for image and video annotation for cultural heritage. She is currently Director of the Italian CINI lab in Artificial Intelligence and Intelligent systems; member of IAPR governing board, recipient of the 2018 Maria Petrou Award, member of the CVF Advisory Board . Next general Chair of ICPR2020, ACM MM2020 and of ECCV 2022. Responsible of 2018-2019 collaboration with Panasonic Beta Lab USA and in 2017 with Facebook FAIR. She is currently working in several Italian and International projects: in H2020 Prystine, Colrobot and Ecsel Arrowhead tool.